Doubts about Sec.7 in Supp. Mat. - Opacity maps visualization

[rugragi2cat]

First of all, thank you for your amazing work. I was trying to get the results you showed in Sec. 7 of supp. mat. and Fig. 5. but I have some doubts.

• Where did you get the data you are using here? My version of the dataset is already super clean (with no outliers).
• I tried to visualize the opacity map on some outliers points (generated randomly) but the opacity is not close to zero.
• How did you get the results in column e in Fig.5? did you combine the reconstruction of all views in one?

Thank you!

[ShunyuanZheng]

Hi, thank you for your interest!

1. The outliers exist in the reconstructed point clouds, but not in the dataset. These data are rendered from Thuman. You can try stage 1 to visualize why the noise emerges.
2. How and where do these outliers generate, whether they are added in 2d images, 2d depths, or 3D space? The network tends to predict the transparent Gaussian primitives near the margin areas with drastically changed depth. I don't know whether it can identify the randomly added noises.
3. The subfigure (e) is the final rendering results, which are rendered from the combined two input views.